Lead scoring
Rank leads by the structure your most valuable customers share, not by firmographics.
Lead scoring decides where sales time goes: which leads a senior account executive calls today, which go through the standard process, and which get nurture sequences instead of direct sales time. This page shows how to frame that decision as a prediction over your CRM and transaction tables, how the scores route leads, and how to measure the ranking.
Beyond firmographic scoring
Traditional lead scoring adds points for a profile: “manufacturing company, 50–200 employees, São Paulo = 75 points.” Two companies with that profile can have opposite outcomes: one becomes your most valuable customer, the other churns in three months.
What separates them is usually the structure around them: who they trade with, who owns them, how they’re growing.
The data
Avra RFM scores every lead from your CRM and transaction tables in context, with your past outcomes as labelled history. A typical schema links each lead to its company, the activities logged against it, and the opportunities it opens, and each company to the invoices it pays once it’s a customer. The schema below is illustrative; see Relational data for how the RFM reads yours.
Won opportunities give you conversions, and invoices give you retention and value. The _at and _on columns are event times, and every table also carries its snapshot time, so a backtest sees only what your CRM knew on each date. See Time and history.
| Data | Purpose |
|---|---|
| Historical outcomes | Converted or not, high value or not, retained or churned: your definition of “good” |
| CRM records | Lead source, engagement history, and account attributes |
| Transaction history | Revenue patterns, product adoption, and usage |
For Brazilian companies, Avra GFM embeddings can add context from outside your data, such as a company’s owners and sector, to your own lead model.
The prediction problem
You name the outcome: converted within 30 days, retained after 12 months, or high lifetime value. Each one names who is scored, the outcome, and the window:
| Goal | Who is scored | Outcome | Window |
|---|---|---|---|
| Conversion | Open leads | An opportunity won | 30 days |
| Retention | Leads that just converted | Still a customer | 12 months |
| Value | Open leads | Revenue invoiced, a number | 12 months |
Conversion alone can favor leads that sign and leave: both companies in the earlier example convert. A retention or value target puts the one that becomes a lasting customer first. Revenue is a number rather than a yes-or-no outcome, so it’s measured by its error against what each customer actually paid. See What you can predict.
From score to routing
Scores rank leads. Set the cutoffs from your own conversion rates by score band.
Use the ranking to route leads. Send the top decile to a senior account executive and fast-track their onboarding. The middle of the ranking goes through your standard sales process, and the bottom gets nurture sequences instead of direct sales time.
Sort the daily lead list by score, so SDRs work top-down, not first in, first out. Start every conversation from an objective baseline: sales adds context instead of starting from zero.
How it’s evaluated
Every model is backtested out of time on your history, against how you rank leads today, and then shadowed on live leads before it routes anything. See Evaluation.
| Metric | What it tells you |
|---|---|
| ROC-AUC | How well scores separate converters from non-converters |
| Lift at the top deciles | How much better the top-ranked leads perform than a random selection, or than your current ranking |
| Conversion by score band | The observed conversion rate in each band, which is how you set routing cutoffs |
| Precision at your team’s capacity | Of the leads your team can work in a day, how many convert |
| Error against realized value | For a revenue target, how far predictions land from what each customer paid |
Integration
| Moment | Path | How it’s used |
|---|---|---|
| The daily lead list | Batch inference | Score every open lead overnight and write the scores to your CRM |
| A new lead arrives | Live predictions | Score it when it’s created, and route it at once |
| Your own lead model | Scores or embeddings as features | Add Avra’s output to the features your model already reads |
Combining with your own model
If you already score leads, add Avra’s output to your features instead of replacing your model:
| Input | Source |
|---|---|
avra_lead_score |
Avra API or batch |
product_engagement |
Your product analytics |
recency |
Your CRM |
source_quality |
Your marketing data |
Or use embeddings as features. Either way, your model gets context your CRM can’t provide.